Transcription
Today, a new AI tool that will complete just about any old task you can throw at it, including QuickBooks stuff, tax stuff? I already don't believe you. I know. And should we even be trusting autonomous AI tools to touch that stuff? No. I last ran a 40-person accounting firm, and I can tell you some of this stuff would have genuinely changed how we work. I am already hearing from real firms who are using it. And by the end of this, I would honestly bet that the majority of you are probably going to find a use case today that you're going to go out and run with.
So today, we are digging into a QuickBooks use case, a tax use case, an Excel use case, a uh, surprise one that's my favorite, and probably most importantly, the rules of the road right now. What I need to communicate to my team because there's a whole lot of stuff we shouldn't be using this for. And that is important. But first, a very quick example of just how this works. I'm in ChatGPT. Where should we begin?
Now, under Tools, you're going to see Agent mode. And it'll give you some suggestions. "Book a dog-friendly Hipcamp with a hut." I don't even know what that is. It can generate reports, perform actions, do spreadsheet stuff, even make presentations. But the easiest thing to do is just, just show you what it does. So, I'll give it a task in QBO. "Head over to QBO and go into every QBO file I have access to. Check each bank connection to see if the bank feed's broken, then report back to me with a table showing any that are..." Tail as old as time. Bank connection breaks, and you only find out about it uh, four hours before when you have to deliver the financials.
I'll enable Agent mode. And this is the magic of Agent. It sets up a little desktop that it's going to work in. So, it's now using its own little computer in this virtual environment. Eventually, it gets to the QBO website. I log it in, and now it is banging around in QBO, jumping between clients like a, like a fresh-checked intern just trying to impress you. This feels so unwise.
A reminder, as we are all fully clenched right now, we're using the ChatGPT Team plan, which does not train the model on any of your prompts. Anything that it sees here. ChatGPT Team, it is the best plan for businesses because of those data protections it comes with. But what if that thing, I don't know, misclicks, hits the wrong button on accident? It can happen. We're going to talk about it.
But look at this, 'cause this is the result. It ended up taking 13 minutes, and it gives us this whole table of all the accounts it has access to and the status of each of those banks. Let me zoom this out. There's, there's so much we're going to have to blur out here. These are all the companies it has access to. On the left, you can see, for example, "Demo Company 3 is like the subscriptions expired." But for all the other companies, it says, "These ones all have active feeds," but this one, it's got two broken feeds. I didn't tell it how to navigate QBO that you have to click the little green icon to go into a company file. Figured all that stuff out on its own. And check this out. If I scroll to the bottom of the result, you'll see this little clock. I can schedule this thing to run on a recurring basis, as much as daily. Pretty, pretty interesting, right?
But for me right now, this use case, it is a no-go. And I'll tell you why. Right now, I don't know how good this is. And I can't trust it to work autonomously. Honestly, if I had like read-only QuickBooks access, maybe. But because this could accidentally click the wrong thing, create an issue, I can't trust it yet. And we're going to get into the rules of the road here at, at the end of the video. But this is what makes this hard: it's an incredibly powerful thing that we don't know that much about yet. It is sometimes incredibly dumb and often breaks. But I'm simultaneously very excited. While I also very much want you to communicate to your team, "Here's the ways I don't want you to use it." Because this is live in ChatGPT Team.
But just some other QuickBooks stuff we've tested. Other ways you can use this, like you've, uh, you wanted to round up sort of a collection of information, like the status of a company file. Give it 8 to 10 things to go out and fetch and report back. Can do that really well. It has no problem navigating the QuickBooks user interface. Have it go into the transaction matching screen and resolve all transactions that meet certain criteria. It can do that. Even pull receipts from all the, the emails that came in yesterday, attach those receipts to transactions, and, and then set that to run every single morning. It's pretty good and it works amazing until it doesn't.
Which brings me to this tax example. Let's say I've got an Excel workbook template. This is an unbelievably cool, free resource. It's like 60 tabs of input data that all rolls up to government forms. It's developed by Glenn Reeves. Gravidatics1040.com. I know a lot of firms that use this to double-check, like, source doc totals to, uh, the output of your tax software. Very useful as is. A lot of people will customize this further for their firm. But let's say this is like a workpaper template that we use. We give this to ChatGPT Agent along with a pile of source docs from a client. So look at this. We've got seven source docs here: brokerage statements, W2s, social security statements, just a hodgepodge of a whole bunch of completed forms I can find online. We are just going to chuck all that stuff into ChatGPT and say, "Here's a spreadsheet that's a workpaper template for my tax firm. Enter all the data from these source documents my client provided into the various tabs in our workpaper template. Don't miss any details. Enter all the items into the spreadsheet." I can like, actually do this, right? Right?
Okay, 15 minutes later. Are you ready? What? Uh, okay. Preview not available. Whoops. Says, "Okay." Okay. Files corrupted. Let's give him another try. You can do it. Just going to restart the same process. Talk amongst yourselves. Okay. Here we go. Take two. It took 14 minutes. It's getting faster. Let's download this. Okay. We got a bunch of values in, uh, the workpaper here. I'm going to have to actually review this. Cue the review montage.
Okay. I, as far as I can tell, it's pulled every detail perfectly. Uh, even the social security statements. Pulled those amounts for the taxpayer, spouse. Pulled everything off the brokerage statement. Actually, it didn't pull the total long-term and short-term transaction summaries because there's not actually an input field for that in the spreadsheet, but you could add your own tab. I like, just think about that. All the values from all seven of those source docs, and this is a 60-tab spreadsheet. It figured out where all that stuff should go, and it did it in 14 minutes. 20, 29 minutes. Okay, took two tries, but the last time when it worked, did it in 14 minutes. Now, that was with seven source docs. There's got to be a ceiling on, like, how much time we'll be willing to work on something. Like, if that's, uh, 50 source docs, then what? Or in that QuickBooks example we did, that was only like six or seven company accounts that I have access to. Now, 'cause I'm, I'm an internet grifter. I'm not, I don't actually run a firm anymore. My old QBOA account had like 350 companies in it. So, will that work on a 350-company file? I think behind the scenes, we've got to remember this Agent is going to have a maximum, like, amount of time that it's allowed to work. Surely that is built into ChatGPT so it just can't run until the end of time. But that, buddy, that's impressive to me. And yes, like, it's slow, it breaks sometimes, but what you're looking at, this is the, like, the new floor for sort of the universal task software can just do for us.
So, does that replace, like, my tax workflow tool? I, right now, probably not. Can it work on my tax software? I know where you're headed. I, I wish it could. Most tax softwares are now desktop-based, and this only works in a browser, but desktop agents are said to be in the works. This sort of thing that works on, like, your entire computer, can access any app that you have on your machine. That stuff's just around the corner. And this use case, uh, for me right now, I've got no problems with this use case. That QBO one, I don't like it 'cause I don't want to have to trust the AI to not get into something that it shouldn't. This, it's only running on a spreadsheet and on the files that I gave it.
And one really interesting thing about ChatGPT Agent, specifically as it applies to our accounting firm workflows, is it's the best AI we've ever had at doing spreadsheet stuff. And that's great for us because virtually all of us have some version of this example. Look at these. These are daily sales reports from a restaurant. This is like a weekly summary. This is a little hand-filled-out thing that somebody in the restaurant did that day. And all that data is supposed to go into this daily summary. I actually heard from an accounting firm owner recently. They work with restaurants and they gather a bunch of these daily reports from their clients and put them into this standardized format. It's a big old pain. But this is now something that ChatGPT Agent is very, very good at.
So what if we give ChatGPT Agent that daily spreadsheet template and then the image that was like the handwritten recap of the day and just say, "Enter the data from the daily sales report image into the daily sales report template spreadsheet." 11 minutes later, it comes back with an Excel file. I download this, and it's exactly what I wanted. It's all of the handwritten data from that day and recap pulled into the spreadsheet template. Isn't that wild? How can we use that? By the way, I didn't have this spreadsheet template. What I did was I took an image from Reddit. I took this image right here. I checked it into a ChatGPT Agent and said, "Give me this in Excel spreadsheet." And here's the result. It took 8 minutes. It not only pulled all the correct values from that example, but it did all the formulas. Look at this. And figured out this is like a total that's a range. This is some sort of calculation. This is the total of those two. Even down here, some of these are kind of ambiguous. What were totals and what weren't? But it worked out the formulas here. It worked out the change bag and deposit were not formulas. Look at this. It's like figuring out the exact cells that went into this formula. That's pretty cool. And most of us have some version of this, like, very client-specific process like this that we're doing on a recurring basis that's really like, just irritating and inane.
This use case for me, like, I'm green lights on this because I don't even have to give it, like, access to a system that it's logged into. The worst it can do is just screw up the task, right? And this, this use case specifically is a great example of how, yes, there's a much more intelligent and automated way to do this, almost certainly. Like, why are they manually filling out this paper sheet? What if we instead had them fill out, like, a web form and it pushes that data to a spreadsheet automatically? For sure, there's a better way to do this entirely. But there's also a reason why that better way hasn't been created yet. Because the people may not know how, because they may not have made the time. And the fascinating thing that I think is about to happen is Agent becomes a dumping ground for all of those low-value things that we now just won't make a more automated version of 'cause Agent can get it done.
And in software development parlance, there's, there's actually a term for this. It's called an abstraction trade-off. And it's when you use, like, a less efficient and more general solution to get a job done that could have been done in a more efficient way. So, for example, for our computers that we use now, back in the day, you had, like, very specific machine code where you write code for, like, every aspect of the hardware in a computer. You manage all the memory and all these things, and you can dial it in so it does a specific job incredibly efficiently. But over time, code languages have gotten more abstracted where they're much more simple and they do stuff like manage memory for you. And so you got a group of people that are like the OG nerds that are like, "No, there's this really more efficient automated way to do this." And then all the other people operate up here and they're like, "Well, this is just faster and easier." ChatGPT Agent is about to do that to a whole bunch of things.
Isn't there a software platform to do that job? Uh, yeah, but I could just chuck it in ChatGPT Agent. Doesn't it take like 20 minutes to run? I just, I just set it to run at 5 every morning. So, while there's a lot of tech that it won't replace, there's a lot of other tech that it could displace simply from it being so easy to use. I often teach software people that they're trying to build solutions to solve like 200 IQ problems when the reality of my day running an accounting firm is my team is weighed down by like these 60 IQ dumb workflow headaches. Uh, the files can't move from this system to that system automatically. We get these reports from clients every day, and we have to enter them into an Excel spreadsheet. That is the stuff that Agent is suited for. And frankly, it can't happen soon enough. It's still pretty rough, but we're starting to see what this future could look like.
Now, I want to show you one other use case that I love that I honestly think you may like, run from here and end up using every single day, and then define, like, what are the rules of the road for our teams. What should we be using this for and not using it for? But first, a word from this video's sponsor. Roll the music. Thanks, Jason. Todd here. It's Ramp. Ramp just launched their own agent, an AI policy agent. Look at this. It compares every expense against your actual company policy and sorts them into those recommended for approval, review, or repayment. Again, it does this for each transaction according to your actual policy. It's like the annoying receipt police you never wanted to have to be. A couple examples for my employees here. Angela Martin spent $712 at Ulta, a restaurant. Approval is recommended, but importantly, it tells you why. It's under the $900 domestic limit and the $150 per person cap. Approved. That's a good bot. Another example, Palin spent $275, but we don't know if there were any other attendees there yet. So, look at this. Finance teams can drill into the policy detail because it depends, was this breakfast and dinner a high-cost city? You can even edit the policy directly in Ramp if, for example, you make exceptions for executives that aren't shared with your team. What very cool new AI stuff from Ramp. I use them myself. They are, they're quite good. Learn more, check out the link in the video description. Back to you.
Okay, so this last example, um, I don't start my day with email. I start my day working on the stuff that I want to get done because email, that's what other people want me to do. But what if, as I start my day, AI could like, just check my unread messages and make sure there isn't anything super spicy that I need to jump into? We've wished AI could do this for a long time. Now, I'd argue it can't. Now, for all the things we've seen so far, you know that you probably don't want this system like logging into your email app every single day. It's just kind of janky, and the notion of it like having the credentials to log in is kind of spooky.
Enter ChatGPT's connectors. Check this out here under Tools. If you go to use connectors, you can add any number of sources to this conversation. If I say, "Connect more," you'll see you can connect Outlook email. You see I've already got Gmail connected. This lets ChatGPT have read-only access to those messages. And it doesn't require, like, pulling up a browser and logging into it. You can even open up access to, like, SharePoint folders, Google Drive folders, which means we can hop in. I'll enable Agent mode, ensure my Gmail is toggled on, and say, "Review my unread emails and give me a short list of what I should know from the most important unread messages." Give it a few minutes to work on it.
And check this out. Here's the output. Let me zoom out. Gives me a big old table of the spiciest stuff I need to be aware of. Pour one out for the editors. There's going to be a lot of pixelating happening here. We've got a, a big bank deposit, an invitation to host a podcast. Down here, it's actually combined two similar messages to say that these people reached out about this, specifically someone with questions about an upcoming event. And if I actually go to my inbox, I've got like 70-something unread messages right now. And it's compressed, like, the spiciest 10 or 12 into an eight-item summary here. Pretty cool, right?
But check this out. Remember what we can do with ChatGPT Agent? Come down here. I can set this to run every morning. So if I clock in at 7:00 a.m. every morning, I can have this run at 6:30. So that rather than going straight to my email inbox and getting distracted, I can take a quick skim of this to see if there's anything I need to handle. If not, I can get down to work, circle back to my email inbox later. That's pretty cool, right? That's pretty darn cool. And honestly, for me, I've got no problem cutting that one loose right now. It only has read access to my email. It's not banging around in some system where I could break something. And the main difference between this and Deep Research right now, because you can do the same thing with Deep Research, is I can't schedule Deep Researches to recur every day or at all. But this, I could set to run every morning.
Now you can see a very near future where there's a whole bunch of tasks, dumb little inane things in our accounting firm. Uh, moving files from here to there, moving the status of projects from this to that, where you're going to have a whole bunch of these things like making you more productive. But where is it at today? And what are the responsible ways that we can use it? As you've heard throughout here, I don't have any problems with security under ChatGPT Team from a first-principle standpoint. And we cover this a lot on this channel. Team comes with all the security that I need: multi-factor authentication. Prompts are never trained into the model. SOC 2 Type 2 security, which is a higher level of security than many of the commonplace apps in the accounting vertical. But the devil's in the details here because I don't yet want to cut an autonomous thing loose inside a system where it could have a whoopsie. And to be clear, I haven't seen any like problematic versions of this yet. It's just still very, very early. Over time, we'll probably build that trust with it.
So right now, very explicitly, what are the rules of the road? Here's what I think you ought to communicate to your team who now has access to this. Number one, don't let it run autonomously anywhere where it has write access. I've heard from, like, um, researchers in academia who love Agent because now you can log it into paywalled research databases and it can work in there. But that's an example of something that is read-only. My QuickBooks account, unless I've got permissions dialed in like, just for the Agent, which is something you could certainly do, there's a risk that it could get into something you don't want it to, right? So, not letting it do its own thing in an app unless it's a read-only environment.
Number two, I'm not going to give it login credentials yet. In fact, if I have to log it into something, I will take control of the browser and log it in myself. That's what we did in the QBO example. It got to the login screen. I took control. I put in the username and password. I did the two-factor thing, and I got it logged in. But I'm not going to tell ChatGPT my credentials and let it do that.
Third thing I'm going to tell my team is, go out to those situations right now. The situations where it doesn't have write access, the tax example, the filling out the restaurant spreadsheet example, the email example, it doesn't have write access. Start testing that stuff. See what works. See what doesn't. See what's today-ready and what's close, where like, maybe with the next update of this thing, it'll be good enough. Because for all the use cases we shouldn't use it for right now, it's still worth exploring what it can do today.
Now, biggest blocker here is going to be security spookies. And this has been, I mean, this has been the case with ChatGPT for years, right? And we've gotten a lot of folks over this where since we've had the Team plan, ChatGPT has been totally fine to use for just about anything, even stuff with client information. And I, I'll, I'll point you to the, the bigger video we've done on this if you're still working through this. But Agent, it's new, comes with its own sort of set of security concerns, stuff that we've got to be thinking about and getting really clear with our team, like, what is inbounds there and what isn't. Because I can cut that thing loose to go and work for 30 minutes doing Lord knows what, which is simultaneously exciting, uh, and nerve-wracking, let's say. Ah, yeah. Because in a world where I've got like unlimited access to free interns, basically AI-powered interns, that's exciting. But imagine running a business where the only thing you had working for you was a thousand interns. I wouldn't, and wish that on my worst enemy.
Now, still getting up to speed with ChatGPT, check out this video. It is the most-watched ChatGPT training for accounting firms on the web. And if you do tax work, check out this one on ChatGPT for research. Going to give you some really cool ideas.
[Music]